research article semantic-based facial image-retrieval...

Research ArticleSemantic-Based Facial Image-Retrieval System withAid of Adaptive Particle Swarm Optimization and SquaredEuclidian Distance

Manikandan Kalimuthu and Ilango Krishnamurthi

Sri Krishna College of Engineering amp Technology Coimbatore 641008 India

Correspondence should be addressed to Manikandan Kalimuthu manikandankprofgmailcom

Received 30 April 2015 Revised 19 July 2015 Accepted 29 July 2015

Academic Editor Antonio Bandera

Copyright copy 2015 M Kalimuthu and I Krishnamurthi This is an open access article distributed under the Creative CommonsAttribution License which permits unrestricted use distribution and reproduction in any medium provided the original work isproperly cited

The semantic-based facial image-retrieval system is concerned with the process of retrieving facial images based on the semanticinformation of query images and database images The image-retrieval systems discussed in the literature have some drawbacksthat degrade the performance of facial image retrieval To reduce the drawbacks in the existing techniques we propose an efficientsemantic-based facial image-retrieval (SFIR) system using APSO and squared Euclidian distance (SED) The proposed techniqueconsists of three stages feature extraction optimization and image retrieval Initially the features are extracted from the databaseimages Low-level features (shape color and texture) and high-level features (face mouth nose left eye and right eye) are the twofeatures used in the feature-extraction process In the second stage a semantic gap between these features is reduced by a well-known adaptive particle swarm optimization (APSO) technique Afterward a squared Euclidian distance (SED) measure will beutilized to retrieve the face images that have less distance with the query imageThe proposed semantic-based facial image-retrieval(SFIR) system with APSO-SED will be implemented in working platform of MATLAB and the performance will be analyzed

1 Introduction

Due to the popularity of digital devices and the rise ofsocial networkphoto sharing services the availability ofconsumer photos is increasing [1 2] A large percentage ofthose photos contain human faces [3] The importance andthe sheer amount of human face photos make the manip-ulation (eg search and mining) of large-scale human faceimages an important research problem and enablemany real-world applications [4] Facial images have gained importanceamongst digital images due to their use in various aspectsof life such as in airports law enforcement applicationssecurity systems and automated surveillance applications[5] Law enforcement and criminal investigation agenciestypically maintain large image databases of human faces [6]Such databases consist of faces of individuals who have eithercommitted crimes or are suspected of having been involved incriminal activities in the past Composite drawings are usedin identifying a potential suspect from an image database [7]Retrieval from these databases is performed in the context of

the following activities the matching of composite drawingsBann file searching and the ranking of a photo lineup [1]

The face-retrieval problem is concerned with retrievingfacial images that are relevant to usersrsquo requests from a col-lection of images Retrieving is based on the visual contentsandor information associated with this facial image [8] Theretrieval system is expected to display the images in thedatabase that match the sketch Bann file searching is per-formedwhen a (suspected) criminal at hand does not disclosehisher legitimate identification information to enable a lawenforcementcriminal investigator to retrieve the criminalrsquospast history [9] Under such circumstances the investigatorvisually scans the criminalrsquos face to extract certain featuresand uses them in performing the retrieval In a ranking of aphoto lineup the person performing the retrieval provides avague and often uncertain set of features of a face and expectsthe system to provide a ranking of faces in the databasethat match the feature descriptions [10] The victim or aneye witness of a crime describes the facial features of theperpetrator of the crime to a forensic composite technician

Hindawi Publishing CorporationJournal of Applied MathematicsVolume 2015 Article ID 284378 12 pageshttpdxdoiorg1011552015284378

2 Journal of Applied Mathematics

[11] The human face is the most significant component ofthe human body that people use for face identification andverification and the recognition of others therefore facialimages are the most common biometric characteristic usedfor human verification and identification [10]

The fivemajor aspects of face study are the representationof faces the detection of faces the identification of facesthe analysis of facial expressions and the classification offaces based on physical features [5] Human attributes (eggender race hairstyle and hair color) are high-level semanticdescriptions of a person Using these human attributesmany researchers have achieved capable results in differentapplications such as face verification identification andkeyword-based face-image retrieval [12 13]The retrieval taskcontains two types of target images One is face images withthe same identity as the query face [14] The other is faceimages that have an appearance similar to the query face [7]The description of a face given by people is almost alwayssemantic in nature using verbal terms such as ldquolong facerdquoldquothick lippedrdquo or ldquoblonde hairedrdquo Semantic face retrievalinvolves the retrieval of face images based on not the rawimage content but the semantics of the facial features suchas the description of a personrsquos nose eyes lips or chinThe semantic face-retrieval system is based on the semanticdescription of the face and then an image match of thecomposed image with the images in the database [15] Face-recognition systems are in high demand in airports and otherpublic places for automated surveillance applications [16]Most of these abovementioned applications use the face asa hard biometric for the verification or identification of aperson and consist primarily of the task of matching theactual image of a face with those stored in a database of faceimages [2] However apart from their use as a hard biometricthe ldquosoftrdquo traits of face modality are used to group peopleinstead of uniquely identifying a person by hisher face Faceimages have been used to identify a personrsquos ethnicity genderand age [6 17] A more interesting application that views theface as a soft biometric is in face-retrieval systems In manylaw enforcement applications the soft biometric [12] traits ofthe face have to be matched to retrieve the target face imagefrom a dataset [18]

The outline of this paper is organized as follows Section 2provides brief descriptions of recent related research Sec-tion 3 reveals the motivation of the work Section 4 presentsour proposed semantic-based facial image retrieval usingAPSO-SED Section 5 shows evaluations of our experimentalresults and discussion and Section 6 presents the paperrsquosconclusions

2 Recent Related Research

Park and Jain [19] proposed a method to utilize demographicinformation and facial marks to improve face-image match-ing and retrieval performance They developed an automaticfacial mark-detection method that uses (1) the active appear-ance model for locating primary facial features (eg eyesnose and mouth) (2) Laplacian-of-Gaussian blob detectionand (3) morphological operators The experimental resultsbased on the FERET database (426 images of 213 subjects)

and two mug-shot databases from the forensic domain (1225images of 671 subjects and 10 000 images of 10 000 subjectsresp) showed that the use of soft biometric traits was ableto improve the face-recognition performance of a state-of-the-art commercial matcher This method does not capturethe semantic aspects of a face Humans by nature tend touse the verbal description of the semantic features (high-level features) to describe what they are looking for andthey encounter difficulty using the language of low-levelfeatures In our system however we propose a new face-image-retrieval technique with semantic face aspects

Thang et al [6] proposed a CBIR that involves thedevelopment of a content-based facial image-retrieval systembased on the constrained independent component analysisOriginating from independent component analysis CICAwas a source-separation technique that uses a priori con-straints to extract the desired independent components (ICs)fromdata By providing query images as the constraints to theCICA the ICs that share similar probabilistic features withthe queries from the database can be extracted Then theseextracted ICs were used to evaluate the rank of each imageaccording to the query The experimental results of theirCBIR system tested with different facial databases showedthat their system improved the retrieval performance byusing a compound query although the retrieved images hadlower accuracy The accuracy of the image retrieval will beimproved in our proposed system

Iqbal et al [20] proposed an image-retrieval techniquefor biometric security purposes which was based on colortexture and shape features controlled by fuzzy heuristicsThe proposed approach was based on the three well-knownalgorithms color histogram texture andmoment invariantsThe color histogram was used to extract the color featuresof an image using four components red green blue andintensity The Gabor filter was used to extract the texturefeatures and the Hu moment invariant was used to extractthe shape features of an image The evaluation was carriedout using the standard precision and recall measures andthe results were compared with the existing approachesThe presented results showed that the proposed approachproduced better results than previous methods The studyused the color and shape features in the image retrieval butthe accuracy was lowThe accuracy of the image retrieval willbe improved in our proposed APSO system

Alattab and Kareem [21] proposed a method for seman-tic feature selection using natural language concepts Theresearch strategized to bridge the semantic gap betweenlow-level image features and high-level semantic conceptsThe semantic features were also integrated directly with theEigen faces and color histogram features for facial imagesearching and retrieval to enhance retrieval accuracy for theuser The Euclidean distance was used for feature classesrsquointegration and classification purposesThe proposed humanfacial image retrieval has been evaluated through severalexperiments using precision and recall methods The resultshave indicated high accuracy which was considered a signif-icant improvement over low-level feature-based facial image-retrieval techniques Although semantic features are used toretrieve images there is a semantic gap in this technique To

Journal of Applied Mathematics 3

fill this semantic gap in the existing technique we proposean innovative method known as adaptive particle swarmoptimization (APSO)

Wang et al [22] proposed an effective Weak Label Reg-ularized Local Coordinate Coding (WLRLCC) techniquewhich exploits the principle of local coordinate coding bylearning sparse features and employs the idea of graph-based weak label regularization to enhance the weak labelsof similar facial images The authors proposed an efficientoptimization algorithm to solve the WLRLCC problemThey conducted extensive empirical studies on several facialimage web databases to evaluate the proposed WLRLCCalgorithm from different aspects The experimental resultsvalidated its efficacy The authors used local coordinatecoding which is a complicated optimization techniqueHence we propose an adaptive particle swarm optimizationalgorithm

3 Problem Definition

Theability to search through images is becoming increasinglyimportant as the number of available images rises drasticallyHowever it is impractical for humans to label every imageavailable and dictate which images are similar to which otherimages Content-based image retrieval (CBIR) is an image-retrieval technique that uses the visual content of an imageto find and retrieve the required images from databasesNormally in CBIR facial images are different from otherimages because facial images are complex multidimensionaland similar in their overall configuration Many techniqueshave been proposed in the field of CBIR with facial imagesfrom face databases In the existing methods image-retrievalsystems primarily use low-level visual features such as colortexture and shape Visual features are extracted automaticallyusing image-processingmethods to represent the raw contentof the image Image retrieval based on color usually yieldsimages with similar colors and image retrieval based onshape yields images that clearly have the same shape [23]Therefore systems that are used for the general purposeof image retrieval using low-level features are not effectivewith facial images particularly when the userrsquos query is averbal description Such descriptions do not capture thesemantic aspects of a face Humans by nature tend touse the verbal description of semantic features (high-levelfeatures) to describe what they are looking for and theyencounter difficulty using the language of low-level featuresOne recently developed CBIR system [19] for biometricsecurity is applied on facial images with low-level featuresThis method acquired approximately 95 accurate face-image retrieval but this method does not apply to real-timedata

Moreover the CBIR based on low-level features doesnot produce accurate results in face images with thesame people at different ages Hence the existing sys-tems present some drawbacks in facial image retrievalwhen the facial images are from different ages that isthey are not considered in the semantic features duringretrieval

4 The Proposed Semantic-Based FacialImage Retrieval Using APSO and SquaredEuclidian Distance (SED)

Wepropose an efficient semantic-based facial image-retrieval(SFIR) system using APSO and squared Euclidian distance(SED)This technique retrieves more related images by usingthe APSO algorithm for the query imagesThe architecture ofour proposed method is shown in Figure 1

Our proposed semantic-based facial image retrieval iscomprised of three stages

(i) Feature extraction

(a) Low-level feature(1) Shape (canny edge detection)(2) Texture (local binary pattern)(3) Color (color histogram equalization)

(b) High-level feature(1) Semantic features (face mouth nose and

left and right eyes)

(ii) Optimization

(a) Adaptive particle swarm optimization

(iii) Image retrieval

(a) Squared Euclidian distance

We considered image database 119868119863 where 119899 is the total number

of images in database 119868119863 Similarly the query image database

is defined as 119876119863 119898 represents the total number of query

images in query database 119876119863 Hence

119868119863= 1198941 1198942 119894

119899

119876119863= 1199021 1199022 119902

119898

(1)

Initially the different-age face images from a larger numberof persons will be collected and stored in the face databaseLet 119868119863= 1198941 1198942 119894

119899 be a face database containing face

images and let 119876119863

= 1199021 1199022 119902

119898 be query database

images After the face images are collected the initial stageis the feature extraction In the feature-extraction stage low-level features such as the shape texture and color and high-level features such as the face mouth nose left eye andright eye will be extracted from the database images andquery database images Afterward based on the featuresobtained the different-age face images will be classified bygenerating rules and semantic labeling will be created forthose face images To obtain better image-retrieval resultsa semantic gap between these database image features andquery database image features will be reduced by a well-known optimization method known as adaptive particleswarm optimization (APSO) Thus we obtain a reduced setof features Afterward a squared Euclidian distance (SED)measure will be utilized to retrieve the face images that havea smaller distance from the query image


Training image

Feature extraction

Low-level features High-level features

Face

Left eye

Nose

Mouth

Right eye

Shape

Texture

Color

detector

LBP

Color histogram

Facial database

Best feature selection

Best features

Euclidean distance

Retrieval of relevant documents

Query image

APSO

Canny edge

Figure 1 Architecture of our proposed semantic-based facial image retrieval


41 Feature Extraction

(i) Low-Level Features

(1) Shape (canny edge detection)(2) Texture (local binary pattern)(3) Color (color histogram equalization)

411 Shape Feature Extraction Using Canny Edge-DetectionTechnique This method effectively detects the edges of thegiven input face image (119894

119899) and the query image (119902

119898)

The edges of the face images are detected using the cannyedge-detection algorithm The process of the edge-detectionalgorithm is given below in detail

Step 1 (smoothing) Any noises present in the image canbe filtered out by the smoothing process with the help of aGaussian filter The Gaussian mask is slid over the image tomanipulate one square of pixels at a time and is smaller thanthe actual image size

Step 2 (finding gradients) From the smoothed image thegradients are found using the Sobel operator In the Sobeloperator a pair of 3 times 3 convolution masks is utilized to findthe gradients in 119909- and 119910-axis directionsThe gradients in the119909-axis direction are given in

119866119883=[[

[

minus1 0 1

minus2 0 2

minus1 0 1

]]

]

(2)

The gradients in 119910-axis direction are given in

119866119884=[[

[

1 2 1

0 0 0

minus1 minus2 minus1

]]

]

(3)

The magnitude of the gradient is approximated using thefollowing equation with the use of (2) and (3) which is alsocalled the edge strength of the gradient

|119866| =1003816100381610038161003816119866119883

1003816100381610038161003816 +1003816100381610038161003816119866119884

1003816100381610038161003816 (4)

The edges are difficult to find their location and the directionsof the edges can be found using

120579 = arctan(1003816100381610038161003816119866119883

10038161003816100381610038161003816100381610038161003816119866119884

1003816100381610038161003816

) (5)

Step 3 (nonmaximum suppression) To obtain the sharpedges from the blurred edges this step is performed byconsidering only the local maxima of the gradients Thegradient direction 120579 is to be rounded nearest to 45∘ by itscorresponding 8-connected neighborhood The magnitudeof the current pixel is compared with the magnitude of thepixel in the positive and negative gradient directions If themagnitude of the current pixel is the greatest value meansthen only that pixelmagnitudersquos value is preserved otherwisethe particular pixel magnitude value is suppressed

110 94 104

98 81 79

73 56 66

1 1 1

1 0

1 0 011100001

255

Image

Figure 2 Architecture of local binary pattern

Step 4 (thresholding) The edge pixel that remains afterthe nonmaximum suppression process is subjected to thethresholding process by choosing the thresholds to find theonly true edges in the imageThe edge pixels that are strongerthan the high threshold are considered strong edge pixels andthe edge pixels that are weaker than the low threshold aresuppressedThe edge pixels between these two thresholds aretaken as weak edge pixels

Step 5 (edge tracking) Edge pixels are divided into connectedBLOBs (Binary Large Objects) using an 8-connected neigh-borhoodThe BLOBs that have at least one of the strong edgepixels are preserved and the other BLOBs without strongedge pixels are suppressed

Thus from the face image (119894119899) we obtain the shape

features of an image sh(119894119899) sh(119902

119898)

412 Texture Feature Extraction Using LBP

(1) Local Binary Pattern The LBP originally appeared asa generic descriptor The operator assigns a label to eachpixel of an image by thresholding a 3 times 3 neighborhoodwith the center pixel value and considering the result as abinary numberThe resulting binary values are read clockwisestarting from the top-left neighbor as shown in Figure 2

Given a pixel position (119909119888 119910119888) LBP is defined as an

ordered set of binary comparisons of pixel intensities betweenthe central pixel and its surrounding pixels The resulting 8bits can be expressed as follows

LBP (119909119888 119910119888) =

7

sum

119899=0

119904 (119897119899minus 119897119888) 2119899 (6)

where 119897119888corresponds to the grey value of the central pixel

(119909119888 119910119888) and 119897

119899corresponds to the grey value of the surround-

ing pixelsThus using the local binary pattern texture feature tx(119894

119899)

tx(119902119898) has been extracted from the face image (119894

119899)


413 Color Feature Extraction Using Histogram

(1) Color FeaturesColor is one of the most widely used visualfeatures in image retrieval [24] Here color features 119888(119894

119899) and

119888(119902119898) are extracted by applying a histogram on the query

image (119902119898) and database image (119894

119899) A color histogram is

a representation of the distribution of colors in an imageFor digital images a color histogram represents the numberof pixels that have colors in each fixed list of color rangesthat span the imagersquos color space in the set of all possiblecolors The histogram provides compact summarization ofthe distribution of data in an image The color histogramis a statistic that can be viewed as an approximation of anunderlying continuous distribution of color values

(2) Color HistogramColor histogram ℎ((119894119899) (119902119898)) = (ℎ

119896((119894119899)

(119902119898)) 119896 = 1 119870) is a119870-dimensional vector such that each

component ℎ119896((119894119899) (119902119898)) represents the relative number of

pixels of color 119862119896in the image that is the fraction of pixels

that are most similar to the corresponding color To buildthe color histogram the image colors should be transformedto an appropriate color space and quantized according to aparticular codebook of size119870

119901(119909)119894 =

119899119894

119899 0 le 119894 lt 119866 (7)

In (7) 119866 being the total number of gray levels in the image119899119894is the number of occurrences of gray level 119894 119899 is the

total number of pixels in the image and 119901(119909)119894 is the image

histogram for pixel value 119894Computationally the color histogram is formed by dis-

crediting the colors within an image and counting thenumber of pixels of each colorThe color descriptors of videocan be global and local Global descriptors specify the overallcolor content of the image but contain no information aboutthe spatial distribution of these colors Local descriptorsrelate to particular image regions and in conjunction withthe geometric properties of the latter also describe the spatialarrangement of the colors By comparing the histogramsignatures of two images and matching the color content ofone image with another the color features 119888(119894

119899) and 119888(119902

119898) can

be extracted

(ii) High-Level FeaturesThemost natural way people describea face is by semantically describing facial features Semanticface retrieval refers to retrieval of face images based on notthe raw image content but the semantics of the facial featuressuch as the description features of the face mouth nose andleft and right eyes Here the high-level semantic features areextracted based on the pixel intensity value which rangesaccording to the image taken Here we use the Viola-Jonesalgorithm for the extraction of high-level features

The Viola-Jones algorithm utilizes a multiscale multi-stage classifier that functions on image intensity informationGenerally this approach scrolls a window across the imageand assigns a binary classifier that discerns between featuresof the face mouth nose and left and right eyes and thebackground This classifier is disciplined with a boostingmachine learning meta-algorithm It is credited as the fastest

A B

C D

1 2

3 4

Figure 3The sum of the pixels within rectangle119863 can be computedwith four array references

and most accurate method for faces in monocular grey-levelimages Viola and Jones developed a real-time face detectorcomprised of a cascade of classifiers trained by AdaBoostEach classifier exercises an integral image filter which isreminiscent of Haar basis functions and can be processedvery fast at any location and scale This is essential to speedup the detector At every level in the cascade a subset offeatures is chosen using a feature selection procedure basedon AdaBoost

The method operates on so-called integral images Eachimage element contains the sum of all pixels values to itsupper left allowing for the constant-time summation ofarbitrary rectangular areas

Integral Image For the original image 119894(119909 119910) the integralimage is defined as follows

119894119894 (119909 119910) = sum

1199091015840le119909

sum

1199101015840le119910

119894 (1199091015840 1199101015840) (8)

Using the following pair recurrence

119904 (119909 119910) = 119904 (119909 119910 minus 1) + 119894 (119909 119910)

119894119894 (119909 119910) = 119894119894 (119909 minus 1 119910) + 119904 (119909 119910)

(9)

(where 119904(119909 119910) is the cumulative row sum 119904(119909 minus1) = 0 plus119894119894(minus1 119910) = 0) the integral image can be processed in onepass over the original image Using the integral image anyrectangular sum can be computed in four array references(see Figure 3)

The value of the integral image at location 1 is the sum ofthe pixels in rectangle 119860 The value at location 2 is 119860 + 119861 atlocation 3 is119860+119862 and at location 4 is119860+119861+119862+119863The sumwithin 119863 can be computed as 4 + 1 minus (2 + 3) Viola-Jonesrsquomodified AdaBoost algorithm is presented in pseudocode[22] below

Viola-Jonesrsquo algorithm is as follows

(i) Given sample images (1199091 1199101) (119909

119899 119910119899) where119910

1=

0 1 for negative and positive examples carry out thefollowing

(ii) Initialize weights 1199081119894

= 12119898 12119897 for 1199101= 0 1

where 119898 and 119897 are the numbers of positive andnegative examples

(iii) For 119905 = 1 119879

(1) normalize the weights 119908119905119894larr 119908119905119894sum119899

119895=1119908119905119895


(2) choose the best weak classifier in regard to theweighted error

120576119905= min119891119901120579

sum

119894

119908119894

1003816100381610038161003816ℎ (119909119894 119891 119901 120579) minus 1199101198941003816100381610038161003816 (10)

(3) define ℎ119905(119909) = ℎ(119909 119891

119905 119901119905 120579119905) where 119891

119905 119901119905 and

120579119905are the minimizers of 120576

119894

(4) update the weights

119908119905+1119894

= 1199081199051198941205731minus119890119894 (11)

where 119890119894= 0 if example 119909

119894is classified correctly

and 119890119894= 1 otherwise and 120573

119905= 120576119905(1 minus 120576

119905)

(iv) The final strong classifier is

119862 (119909) =

1 if119879

sum

119905=1

120572119905ℎ119905 (119909) ge

1

2

119879

sum

119905=1

120572119905

0 otherwise(12)

where 120572119905= log(1120573

119905)

Thus the semantic features of the database face images 119891(119894119899)

119898(119894119899) 119899(119894119899) 119897(119894119899) and 119903(119894

119899) and query images 119891(119902

119898) 119898(119902

119898)

119899(119902119898) 119897(119902119898) and 119903(119902

119898) are extracted based on the pixel

intensity value

42 Optimization To obtain better image-retrieval resultsthe feature set of the query dataset will be reduced by thewell-known optimization method of the computer-visionplatform adaptive particle swarm optimization (APSO)Particle swarm optimization (PSO) is a population-basedsearch algorithm It is created to imitate the behavior of birdsin search for food on a cornfield or a fish school The methodcan efficiently find optimal or near-optimal solutions in largesearch spaces

Each individual particle 119894 has a randomly initializedposition 119875

119894= (1199011

119894 1199012

119894 119901

119889

119894) where 119901119889

119894is its position in the

119889th dimension The velocity 119881119894= (V1119894 V2119894 V119889

119894) where V119889

119894is

the velocity in the 119889th dimension 119887119901119901119894= (1198871199011

119894 1198871199012

119894 119887119901

119889

119894)

where 119887119901119889119894is the best position in the 119889th dimension and 119892119901 =

(1198921199011 1198921199012 119892119901

119889) where 119892119901119889 is the global best position in

the 119889th dimension in the 119863-dimensional search space Anyparticle canmove in the direction of its personal best positionto its best global position in the course of each generationThemoving process of a swarm particle in the search space isdescribed as

119881119889

119894= 119881119889

119894+ 1198881sdot 1199031sdot (119887119901119889

119894minus 119901119889

119894) + 1198882sdot 1199032sdot (119892119901119889minus 119901119889

119894) (13)

119901119889

119894= 119901119889

119894+ 120575119881119889

119894 (14)

In (13)

1198881 1198882are constants with the value of 20

1199031 1199032are independent random numbers generated in

the range [01]

Generation of population

Fitness calculation

Compute acceleration factors and velocity

Terminationreached

Updating swarm

Stop

Yes

No

Initialize gb and Pb

Figure 4 Structural design of adaptive particle swarmoptimization

119881119889

119894is the velocity of the 119894th particle

119901119889

119894is the current position of particle 119894

119887119901119889

119894is the best fitness value of the particle at the

current iteration119892119901119889 is the best fitness value in the swarm

The adaptive particle swarm optimization (APSO) methodis used here It provides more accurate results The APSOacceleration coefficients are determined by

1198991198881=2

3(1198881max minus 1198881min) (

119891min119891avg

+119891min2119891max

) + 1198881min

1198991198882=2

3(1198882max minus 1198882min) (

119891min119891avg

+119891min2119891max

) + 1198882min

(15)

where 1198881max 1198881min are minimum and maximum values of

1198881 119891min 119891avg 119891max are minimum average and maximum

fitness value of the particles and 1198882max 1198882min are minimum

and maximum values of 1198882

421 APSO in terms of Parameter Optimization PhasesFigure 4 shows the structural design of adaptive particleswarm optimization to identify the reduced set of features for


image retrieval The steps for parameter optimization are asfollows

(i) Generate the particle randomly for population size119898generate the particles randomly

(ii) Describe the fitness function choose the fitness func-tion which should be used for the constraints accord-ing to the current population Here (16) is used tocalculate fitness function

fitness = min (119901119889) (119888) max (119897119904) (119887119908) (119901119904) (16)

(iii) Initialize 119892119901 and 119887119901 initially the fitness value iscalculated for each particle and set as the Pbest valueof each particleThe best Pbest value is selected as the119892119901 value

(iv) Calculations of the acceleration factors the accelera-tion factors are computed using (15)

(v) Velocity computation the new velocity is calculatedusing the equation below Substituting 119899119888

1and 119899119888

2

values in the velocity equation (8) the equationbecomes

119881119889

119894= 119908119889119881119889

119894+ 1198991198881sdot 1199031sdot (119887119901119889

119894minus 119901119889

119894) + 119899119888

2sdot 1199032

sdot (119892119901119889minus 119901119889

119894)

(17)

(vi) Swarmupdate calculate the fitness function again andupdate 119887119901 and 119892119901 values If the new value is betterthan the previous one replace the oldwith the currentone and select the best 119887119901 as 119892119901

(vii) Criterion to stop continue until the solution is goodenough or a maximum iteration is reached

Thus we obtained a reduced set of features from the APSOtechnique This reduced set of features will be then subjectedto image retrieval using the squared minimum distancecalculation

43 Image Retrieval

431 Distance Measurement Distance measurement is animportant stage in image retrieval A query image is given to asystem that retrieves a similar image from the image databaseOur main objective is to retrieve images by measuring thedistance between the reduced set of query image featuresthat is obtained from the optimization process and the imagefeatures in the database To retrieve the image a squaredminimum distance measure is computed The process of thesquared minimum distance measure is described below

dist (119909 119910) (119886 119887) = radic(119909 minus 119886)2 + (119910 minus 119887)2 (18)

Here (119909 119910) are the database images and (119886 119887) are the queryimages

By exploiting (18) the relevant images are extractedwhich have the minimum value of SED By following theaforementioned process images similar to the query imagesare successfully retrieved from the image database

5 Experimental Results

The proposed facial image-retrieval technique using APSO isimplemented in the working platform of MATLAB (version713) with a machine configuration as follows

Processor Intel Core i3OS Windows 7XPCPU speed 320GHzRAM 4GB

The proposed facial image-retrieval technique with the aid ofAPSO is analyzed using the Adience dataset [25] the IMMface dataset [26] and the locally collected facial dataset Thefacial database consists of a total of 961 images which includechildren middle-aged people and old people For training313 images were used 641 images were used for the testingprocess

51 Performance Analysis The performance of the proposedtechnique is compared with PSO and GA algorithms to showthe efficiency of the proposed technique The performance isevaluated by three quantitative performance metrics

(i) 119865-measure(ii) Precision(iii) Recall

511 Precision and Recall Values Precision is the fraction ofretrieved images that are relevant to the findings and recallin retrieval is the fraction of images relevant to the query thatare successfully retrieved

The precision and recall values are calculated using thefollowing equations

119901 =119873119903119903

119873119903

119877 =119873119903119903

119873119889

(19)

(i) 119873119903119903denotes the number of relevant images retrieved

(ii) 119873119903represents the total of retrieved images

(iii) 119873119889is the total number of images in the database

512 119865-Measure A measure that combines precision andrecall is the traditional 119865-measure

119865 = 2(Precision sdot RecallPrecision + Recall

) (20)

Figures 5 6 and 7 represent the sample input training imagesof children old people andmiddle-aged people respectively

From the input images high-level features and low-levelfeatures are extracted and the extracted features are given toAPSO to select the optimal features After that the Euclideandistance is calculated between the features of the query image


Figure 5 Input images of children

Figure 6 Input images of old people

Figure 7 Input images of middle-aged people

Table 1 Precision recall and 119865-measure values for children

Measurements Proposed technique PSO GAPrecision 09665 09402 09326Recall 1 1 1119865-measure 04871 04811 0480

and the best features Finally the images that are similar to thedistance are retrieved

Figure 8(a) shows the query images of children and thecorresponding retrieved images are given in Figure 8(b)

Table 1 shows the performance measures such as preci-sion recall and 119865-measures calculated using the statisticalmeasures given in [27] for children

Precision is one of the important performance measuresin the image-retrieval process Table 1 shows that the pre-cision rate of the proposed technique is higher than that

of the other techniques Similarly the 119865-measure is higherthan the other techniques Although the difference is smallit indicates the improvement in the performance of theproposed technique

Figure 9(a) shows the query images of middle-aged peo-ple and the corresponding retrieved images are given in Fig-ure 9(b) In Table 2 performance measures such as precisionrecall and 119865-measure are given for the proposed techniquePSO and GA in middle-aged people image retrieval Theperformance measures are higher than the other techniquesexcept for recall which implies that the proposed techniqueperforms better than the other techniques

Figure 10(a) shows the query images of old people andthe corresponding retrieved images are given in Figure 10(b)In Table 3 the performance of the proposed technique is ana-lyzed in old people image retrieval using precision recall and119865-measure and these performance measures are comparedwith other techniques such as PSO and GA Compared to


(a) (b)

Figure 8 (a) Query images of children and (b) retrieved images

(a) (b)

Figure 9 (a) Query images of middle-aged people and (b) retrieved images

(a) (b)

Figure 10 (a) Query images of old people and (b) retrieved images

the other techniques (PSO and GA) the proposed techniquehas a higher rate of precision and 119865-measure

Discussion The average rate for all measures is calculatedfrom Tables 1 2 and 3 and a graph is drawn and given

in Figure 11 The precision rate of the proposed techniqueis 09684 and the other techniques such as PSO and GAhave precision rates of 09399 and 09248 respectively Theproposed technique has a higher precision rate than othertechniques such as PSO and GA The variation in the


0010203040506070809

1

Precision RecallProposed technique 09684 1 04921PSO 09399 1 04872GA 09248 1 04821

F-measure

Figure 11 Performance measures of the proposed technique PSOand GA

Table 2 Precision recall and 119865-measure values for middle-agedpeople


Table 3 Precision recall and 119865-measure values for old people


precision rate between the proposed technique and the othertechniques is 00285ndash0436 Furthermore the performance ofthe proposed technique is evaluated using the119865-measureThe119865-measures of PSO and GA are 04872 and 04821 respec-tively The 119865-measure of the proposed technique is 04921which is 00490 to 00101 greater than the other techniquesTherefore the proposed APSO-based facial image-retrievaltechnique performs better than other techniques

6 Conclusions

In this paper we propose an efficient semantic-based facialimage-retrieval (SFIR) system using APSO and the squaredEuclidian distance (SED)method to reduce the shortcomingsof the existing methods Moreover in comparative analysisour proposed technique performance is compared with exist-ing methods The proposed technique has a higher precisionrate and 119865-measure than other techniques such as PSO andGA The comparison result shows that our proposed APSO-based semantic-based facial image-retrieval (SFIR) techniqueretrieved the images more accurately than existing methodsHence it is shown that our proposed semantic-based facialimage-retrieval (SFIR) system using the APSO techniquemore precisely and efficiently retrieves images by achievinga higher retrieval rate

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] G F Ahmed and R Barskar ldquoA study on different imageretrieval techniques in image processingrdquo International Journalof Soft Computing and Engineering vol 1 no 4 pp 247ndash2512011

[2] P Kannan P S Bala andG Aghila ldquoMultimedia retrieval usingontology for semantic webmdasha surveyrdquo International Journal ofSoft Computing and Engineering vol 2 no 1 pp 47ndash51 2012

[3] M Bhagat ldquoFace image retrieval using sparse code words withencryptionrdquo Sinhgad Institute of Management and ComputerApplication vol 1 no 1 pp 180ndash183 2011

[4] Y Suzuki and T Shibata ldquoIllumination-invariant face identifi-cation using edge-based feature vectors in pseudo-2D HiddenMarkovModelsrdquo in Proceedings of the 14th European Signal Pro-cessing Conference (EUSIPCO rsquo06) Florence Italy September2006

[5] V Gudivada V Raghavan and G Seetharaman ldquoAn approachto interactive retrieval in face image databases based on seman-tic attributesrdquo International Journal of Artificial Intelligence ampApplications vol 2 no 3 2011

[6] N D Thang T Rasheed Y-K Lee S Lee and T-S KimldquoContent-based facial image retrieval using constrained inde-pendent component analysisrdquo Information Sciences vol 181 no15 pp 3162ndash3174 2011

[7] V Ganesh and J M E Madhavan ldquoFace image retrievalusing Lwt-Pca with different classifiersrdquo International Journal ofEngineering and Computer Science vol 3 no 3 pp 5105ndash51082014

[8] H Kaur and K Jyoti ldquoSurvey of techniques of high level seman-tic based image retrievalrdquo International Journal of Research inComputer and Communication Technology vol 2 no 1 pp 15ndash19 2013

[9] V Vijayarajan M Khalid and C Mouli ldquoA review from key-word based image retrieval to ontology based image retrievalrdquoInternational Journal of Reviews in Computing vol 12 no 1 pp1ndash17 2012

[10] B-C Chen Y-Y Chen Y-HKuo andWHHsu ldquoScalable faceimage retrieval using attribute-enhanced sparse codewordsrdquoIEEE Transactions on Multimedia vol 15 no 5 pp 1163ndash11732013

[11] J Yang M Luo and Y Jiao ldquoFace recognition based on imagelatent semantic analysis model and SVMrdquo International Journalof Signal Processing Image Processing and Pattern Recognitionvol 6 no 3 pp 101ndash110 2013

[12] I Sudha V SaradhaM Tamilselvi andD Vennila ldquoFace imageretrieval using facial attributes By K-meansrdquo InternationalJournal of Computer Science and Information Technologies vol5 no 2 pp 1622ndash1625 2014

[13] A Dinakaran and A Kumar ldquoInteractive image retrieval usingtext and image contentrdquo Cybernetics and Information Technolo-gies vol 10 no 3 pp 20ndash30 2010

[14] Petchimuthu and Benifa ldquoFace image retrieval using attributesparse codewordrsquosrdquo International Journal of Advanced Researchin Computer Science and Software Engineering vol 4 no 3 pp124ndash127 2014


[15] S Pange and S Lokhande ldquoImage retrieval system by usingCWT and support vector machinesrdquo Signal amp Image Processingvol 3 no 3 pp 63ndash71 2012

[16] D Depalov T Pappas D Li and B Gandhi ldquoA perceptualapproach for semantic image retrievalrdquo American Journal ofScientific Research vol 1 no 11 pp 6ndash19 2010

[17] V Huddy S R Schweinberger I Jentzsch and A M BurtonldquoMatching faces for semantic information and names an event-related brain potentials studyrdquo Cognitive Brain Research vol 17no 2 pp 314ndash326 2003

[18] M Balaganesh and N Arthi ldquoImage retrieval using attributeenhanced sparse code wordsrdquo International Journal of Innova-tive Research in Science Engineering and Technology vol 3 no5 pp 12580ndash12588 2014

[19] U Park and A K Jain ldquoFace matching and retrieval using softbiometricsrdquo IEEE Transactions on Information Forensics andSecurity vol 5 no 3 pp 406ndash415 2010

[20] K Iqbal M O Odetayo and A James ldquoContent-based imageretrieval approach for biometric security using colour textureand shape features controlled by fuzzy heuristicsrdquo Journal ofComputer and SystemSciences vol 78 no 4 pp 1258ndash1277 2012

[21] AAAlattab and SAKareem ldquoSemantic features selection andrepresentation for facial image retrieval systemrdquo in Proceedingsof the 4th International Conference on Intelligent Systems Mod-elling and Simulation (ISMS rsquo13) pp 299ndash304 January 2013

[22] D Wang S Hoi Y He J Zhu T Mei and J Luo ldquoRetrieval-based face annotation byweak label regularized local coordinatecodingrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 36 no 3 pp 550ndash563 2014

[23] G Reddy G R Babu and P V Somasekhar ldquoImage retrievalby semantic indexingrdquo Journal of Theoretical and AppliedInformation Technology vol 2 no 1 pp 745ndash750 2010

[24] M Rehman M Iqbal M Sharif and M Raza ldquoContent basedface image retrieval surveyrdquoWorldApplied Sciences Journal vol19 no 3 pp 404ndash412 2012

[25] httpwwwopenuacilhomehassnerAdiencedatahtml[26] httpwww2immdtudkpubdbviewspublication details

phpid=3160[27] httpsenwikipediaorgwikiSensitivity and specificit

Submit your manuscripts athttpwwwhindawicom

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

MathematicsJournal of


Mathematical Problems in Engineering

Hindawi Publishing Corporationhttpwwwhindawicom

Differential EquationsInternational Journal of

Volume 2014

Applied MathematicsJournal of


Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of


Mathematical PhysicsAdvances in

Complex AnalysisJournal of


OptimizationJournal of


CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of


Operations ResearchAdvances in

Journal of


Function Spaces

Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of Mathematics and Mathematical Sciences


The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Algebra

Discrete Dynamics in Nature and Society



Decision SciencesAdvances in

Discrete MathematicsJournal of


Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Stochastic AnalysisInternational Journal of


[11] The human face is the most significant component ofthe human body that people use for face identification andverification and the recognition of others therefore facialimages are the most common biometric characteristic usedfor human verification and identification [10]

The fivemajor aspects of face study are the representationof faces the detection of faces the identification of facesthe analysis of facial expressions and the classification offaces based on physical features [5] Human attributes (eggender race hairstyle and hair color) are high-level semanticdescriptions of a person Using these human attributesmany researchers have achieved capable results in differentapplications such as face verification identification andkeyword-based face-image retrieval [12 13]The retrieval taskcontains two types of target images One is face images withthe same identity as the query face [14] The other is faceimages that have an appearance similar to the query face [7]The description of a face given by people is almost alwayssemantic in nature using verbal terms such as ldquolong facerdquoldquothick lippedrdquo or ldquoblonde hairedrdquo Semantic face retrievalinvolves the retrieval of face images based on not the rawimage content but the semantics of the facial features suchas the description of a personrsquos nose eyes lips or chinThe semantic face-retrieval system is based on the semanticdescription of the face and then an image match of thecomposed image with the images in the database [15] Face-recognition systems are in high demand in airports and otherpublic places for automated surveillance applications [16]Most of these abovementioned applications use the face asa hard biometric for the verification or identification of aperson and consist primarily of the task of matching theactual image of a face with those stored in a database of faceimages [2] However apart from their use as a hard biometricthe ldquosoftrdquo traits of face modality are used to group peopleinstead of uniquely identifying a person by hisher face Faceimages have been used to identify a personrsquos ethnicity genderand age [6 17] A more interesting application that views theface as a soft biometric is in face-retrieval systems In manylaw enforcement applications the soft biometric [12] traits ofthe face have to be matched to retrieve the target face imagefrom a dataset [18]

The outline of this paper is organized as follows Section 2provides brief descriptions of recent related research Sec-tion 3 reveals the motivation of the work Section 4 presentsour proposed semantic-based facial image retrieval usingAPSO-SED Section 5 shows evaluations of our experimentalresults and discussion and Section 6 presents the paperrsquosconclusions

2 Recent Related Research

Park and Jain [19] proposed a method to utilize demographicinformation and facial marks to improve face-image match-ing and retrieval performance They developed an automaticfacial mark-detection method that uses (1) the active appear-ance model for locating primary facial features (eg eyesnose and mouth) (2) Laplacian-of-Gaussian blob detectionand (3) morphological operators The experimental resultsbased on the FERET database (426 images of 213 subjects)

and two mug-shot databases from the forensic domain (1225images of 671 subjects and 10 000 images of 10 000 subjectsresp) showed that the use of soft biometric traits was ableto improve the face-recognition performance of a state-of-the-art commercial matcher This method does not capturethe semantic aspects of a face Humans by nature tend touse the verbal description of the semantic features (high-level features) to describe what they are looking for andthey encounter difficulty using the language of low-levelfeatures In our system however we propose a new face-image-retrieval technique with semantic face aspects

Thang et al [6] proposed a CBIR that involves thedevelopment of a content-based facial image-retrieval systembased on the constrained independent component analysisOriginating from independent component analysis CICAwas a source-separation technique that uses a priori con-straints to extract the desired independent components (ICs)fromdata By providing query images as the constraints to theCICA the ICs that share similar probabilistic features withthe queries from the database can be extracted Then theseextracted ICs were used to evaluate the rank of each imageaccording to the query The experimental results of theirCBIR system tested with different facial databases showedthat their system improved the retrieval performance byusing a compound query although the retrieved images hadlower accuracy The accuracy of the image retrieval will beimproved in our proposed system

Iqbal et al [20] proposed an image-retrieval techniquefor biometric security purposes which was based on colortexture and shape features controlled by fuzzy heuristicsThe proposed approach was based on the three well-knownalgorithms color histogram texture andmoment invariantsThe color histogram was used to extract the color featuresof an image using four components red green blue andintensity The Gabor filter was used to extract the texturefeatures and the Hu moment invariant was used to extractthe shape features of an image The evaluation was carriedout using the standard precision and recall measures andthe results were compared with the existing approachesThe presented results showed that the proposed approachproduced better results than previous methods The studyused the color and shape features in the image retrieval butthe accuracy was lowThe accuracy of the image retrieval willbe improved in our proposed APSO system

Alattab and Kareem [21] proposed a method for seman-tic feature selection using natural language concepts Theresearch strategized to bridge the semantic gap betweenlow-level image features and high-level semantic conceptsThe semantic features were also integrated directly with theEigen faces and color histogram features for facial imagesearching and retrieval to enhance retrieval accuracy for theuser The Euclidean distance was used for feature classesrsquointegration and classification purposesThe proposed humanfacial image retrieval has been evaluated through severalexperiments using precision and recall methods The resultshave indicated high accuracy which was considered a signif-icant improvement over low-level feature-based facial image-retrieval techniques Although semantic features are used toretrieve images there is a semantic gap in this technique To














(ii) Optimization








119868119863= 1198941 1198942 119894

119899

119876119863= 1199021 1199022 119902

119898

(1)




= 1199021 1199022 119902




Training image

Feature extraction


Face

Left eye

Nose

Mouth

Right eye

Shape

Texture

Color

detector

LBP

Color histogram

Facial database


Best features

Euclidean distance


Query image

APSO

Canny edge








119898)




119866119883=[[

[

minus1 0 1

minus2 0 2

minus1 0 1

]]

]

(2)


119866119884=[[

[

1 2 1

0 0 0


]]

]

(3)


|119866| =1003816100381610038161003816119866119883

1003816100381610038161003816 +1003816100381610038161003816119866119884

1003816100381610038161003816 (4)


120579 = arctan(1003816100381610038161003816119866119883

10038161003816100381610038161003816100381610038161003816119866119884

1003816100381610038161003816

) (5)


110 94 104

98 81 79

73 56 66

1 1 1

1 0

1 0 011100001

255

Image






119898)





LBP (119909119888 119910119888) =

7

sum

119899=0

119904 (119897119899minus 119897119888) 2119899 (6)


(119909119888 119910119888) and 119897



119899)


119899)




119899) and






119896((119894119899)





119901(119909)119894 =

119899119894

119899 0 le 119894 lt 119866 (7)





119899) and 119888(119902

119898) can

be extracted



A B

C D

1 2

3 4





119894119894 (119909 119910) = sum

1199091015840le119909

sum

1199101015840le119910

119894 (1199091015840 1199101015840) (8)


119904 (119909 119910) = 119904 (119909 119910 minus 1) + 119894 (119909 119910)

119894119894 (119909 119910) = 119894119894 (119909 minus 1 119910) + 119904 (119909 119910)

(9)





119899 119910119899) where119910

1=



= 12119898 12119897 for 1199101= 0 1


(iii) For 119905 = 1 119879


119895=1119908119905119895



120576119905= min119891119901120579

sum

119894

119908119894

1003816100381610038161003816ℎ (119909119894 119891 119901 120579) minus 1199101198941003816100381610038161003816 (10)

(3) define ℎ119905(119909) = ℎ(119909 119891

119905 119901119905 120579119905) where 119891

119905 119901119905 and


119894


119908119905+1119894

= 1199081199051198941205731minus119890119894 (11)

where 119890119894= 0 if example 119909



119905= 120576119905(1 minus 120576

119905)


119862 (119909) =

1 if119879

sum

119905=1

120572119905ℎ119905 (119909) ge

1

2

119879

sum

119905=1

120572119905

0 otherwise(12)

where 120572119905= log(1120573

119905)


119898(119894119899) 119899(119894119899) 119897(119894119899) and 119903(119894

119899) and query images 119891(119902

119898) 119898(119902

119898)

119899(119902119898) 119897(119902119898) and 119903(119902


intensity value



119894= (1199011

119894 1199012

119894 119901

119889

119894) where 119901119889



119894) where V119889

119894is


119894 1198871199012

119894 119887119901

119889

119894)


(1198921199011 1198921199012 119892119901



119881119889

119894= 119881119889

119894+ 1198881sdot 1199031sdot (119887119901119889

119894minus 119901119889

119894) + 1198882sdot 1199032sdot (119892119901119889minus 119901119889

119894) (13)

119901119889

119894= 119901119889

119894+ 120575119881119889

119894 (14)

In (13)



the range [01]


Fitness calculation


Terminationreached

Updating swarm

Stop

Yes

No



119881119889


119901119889


119887119901119889




1198991198881=2

3(1198881max minus 1198881min) (

119891min119891avg

+119891min2119891max

) + 1198881min

1198991198882=2

3(1198882max minus 1198882min) (

119891min119891avg

+119891min2119891max

) + 1198882min

(15)










fitness = min (119901119889) (119888) max (119897119904) (119887119908) (119901119904) (16)




1and 119899119888

2


119881119889

119894= 119908119889119881119889

119894+ 1198991198881sdot 1199031sdot (119887119901119889

119894minus 119901119889

119894) + 119899119888

2sdot 1199032

sdot (119892119901119889minus 119901119889

119894)

(17)




43 Image Retrieval













119901 =119873119903119903

119873119903

119877 =119873119903119903

119873119889

(19)






) (20)

















(a) (b)


(a) (b)


(a) (b)






0010203040506070809

1


F-measure







6 Conclusions




References



































Volume 2014




Journal of











Journal of


Function Spaces






Algebra






















(ii) Optimization








119868119863= 1198941 1198942 119894

119899

119876119863= 1199021 1199022 119902

119898

(1)




= 1199021 1199022 119902




Training image

Feature extraction


Face

Left eye

Nose

Mouth

Right eye

Shape

Texture

Color

detector

LBP

Color histogram

Facial database


Best features

Euclidean distance


Query image

APSO

Canny edge








119898)




119866119883=[[

[

minus1 0 1

minus2 0 2

minus1 0 1

]]

]

(2)


119866119884=[[

[

1 2 1

0 0 0


]]

]

(3)


|119866| =1003816100381610038161003816119866119883

1003816100381610038161003816 +1003816100381610038161003816119866119884

1003816100381610038161003816 (4)


120579 = arctan(1003816100381610038161003816119866119883

10038161003816100381610038161003816100381610038161003816119866119884

1003816100381610038161003816

) (5)


110 94 104

98 81 79

73 56 66

1 1 1

1 0

1 0 011100001

255

Image






119898)





LBP (119909119888 119910119888) =

7

sum

119899=0

119904 (119897119899minus 119897119888) 2119899 (6)


(119909119888 119910119888) and 119897



119899)


119899)




119899) and






119896((119894119899)





119901(119909)119894 =

119899119894

119899 0 le 119894 lt 119866 (7)





119899) and 119888(119902

119898) can

be extracted



A B

C D

1 2

3 4





119894119894 (119909 119910) = sum

1199091015840le119909

sum

1199101015840le119910

119894 (1199091015840 1199101015840) (8)


119904 (119909 119910) = 119904 (119909 119910 minus 1) + 119894 (119909 119910)

119894119894 (119909 119910) = 119894119894 (119909 minus 1 119910) + 119904 (119909 119910)

(9)





119899 119910119899) where119910

1=



= 12119898 12119897 for 1199101= 0 1


(iii) For 119905 = 1 119879


119895=1119908119905119895



120576119905= min119891119901120579

sum

119894

119908119894

1003816100381610038161003816ℎ (119909119894 119891 119901 120579) minus 1199101198941003816100381610038161003816 (10)

(3) define ℎ119905(119909) = ℎ(119909 119891

119905 119901119905 120579119905) where 119891

119905 119901119905 and


119894


119908119905+1119894

= 1199081199051198941205731minus119890119894 (11)

where 119890119894= 0 if example 119909



119905= 120576119905(1 minus 120576

119905)


119862 (119909) =

1 if119879

sum

119905=1

120572119905ℎ119905 (119909) ge

1

2

119879

sum

119905=1

120572119905

0 otherwise(12)

where 120572119905= log(1120573

119905)


119898(119894119899) 119899(119894119899) 119897(119894119899) and 119903(119894

119899) and query images 119891(119902

119898) 119898(119902

119898)

119899(119902119898) 119897(119902119898) and 119903(119902


intensity value



119894= (1199011

119894 1199012

119894 119901

119889

119894) where 119901119889



119894) where V119889

119894is


119894 1198871199012

119894 119887119901

119889

119894)


(1198921199011 1198921199012 119892119901



119881119889

119894= 119881119889

119894+ 1198881sdot 1199031sdot (119887119901119889

119894minus 119901119889

119894) + 1198882sdot 1199032sdot (119892119901119889minus 119901119889

119894) (13)

119901119889

119894= 119901119889

119894+ 120575119881119889

119894 (14)

In (13)



the range [01]


Fitness calculation


Terminationreached

Updating swarm

Stop

Yes

No



119881119889


119901119889


119887119901119889




1198991198881=2

3(1198881max minus 1198881min) (

119891min119891avg

+119891min2119891max

) + 1198881min

1198991198882=2

3(1198882max minus 1198882min) (

119891min119891avg

+119891min2119891max

) + 1198882min

(15)










fitness = min (119901119889) (119888) max (119897119904) (119887119908) (119901119904) (16)




1and 119899119888

2


119881119889

119894= 119908119889119881119889

119894+ 1198991198881sdot 1199031sdot (119887119901119889

119894minus 119901119889

119894) + 119899119888

2sdot 1199032

sdot (119892119901119889minus 119901119889

119894)

(17)




43 Image Retrieval













119901 =119873119903119903

119873119903

119877 =119873119903119903

119873119889

(19)






) (20)

















(a) (b)


(a) (b)


(a) (b)






0010203040506070809

1


F-measure







6 Conclusions




References



































Volume 2014




Journal of











Journal of


Function Spaces






Algebra










Training image

Feature extraction


Face

Left eye

Nose

Mouth

Right eye

Shape

Texture

Color

detector

LBP

Color histogram

Facial database


Best features

Euclidean distance


Query image

APSO

Canny edge








119898)




119866119883=[[

[

minus1 0 1

minus2 0 2

minus1 0 1

]]

]

(2)


119866119884=[[

[

1 2 1

0 0 0


]]

]

(3)


|119866| =1003816100381610038161003816119866119883

1003816100381610038161003816 +1003816100381610038161003816119866119884

1003816100381610038161003816 (4)


120579 = arctan(1003816100381610038161003816119866119883

10038161003816100381610038161003816100381610038161003816119866119884

1003816100381610038161003816

) (5)


110 94 104

98 81 79

73 56 66

1 1 1

1 0

1 0 011100001

255

Image






119898)





LBP (119909119888 119910119888) =

7

sum

119899=0

119904 (119897119899minus 119897119888) 2119899 (6)


(119909119888 119910119888) and 119897



119899)


119899)




119899) and






119896((119894119899)





119901(119909)119894 =

119899119894

119899 0 le 119894 lt 119866 (7)





119899) and 119888(119902

119898) can

be extracted



A B

C D

1 2

3 4





119894119894 (119909 119910) = sum

1199091015840le119909

sum

1199101015840le119910

119894 (1199091015840 1199101015840) (8)


119904 (119909 119910) = 119904 (119909 119910 minus 1) + 119894 (119909 119910)

119894119894 (119909 119910) = 119894119894 (119909 minus 1 119910) + 119904 (119909 119910)

(9)





119899 119910119899) where119910

1=



= 12119898 12119897 for 1199101= 0 1


(iii) For 119905 = 1 119879


119895=1119908119905119895



120576119905= min119891119901120579

sum

119894

119908119894

1003816100381610038161003816ℎ (119909119894 119891 119901 120579) minus 1199101198941003816100381610038161003816 (10)

(3) define ℎ119905(119909) = ℎ(119909 119891

119905 119901119905 120579119905) where 119891

119905 119901119905 and


119894


119908119905+1119894

= 1199081199051198941205731minus119890119894 (11)

where 119890119894= 0 if example 119909



119905= 120576119905(1 minus 120576

119905)


119862 (119909) =

1 if119879

sum

119905=1

120572119905ℎ119905 (119909) ge

1

2

119879

sum

119905=1

120572119905

0 otherwise(12)

where 120572119905= log(1120573

119905)


119898(119894119899) 119899(119894119899) 119897(119894119899) and 119903(119894

119899) and query images 119891(119902

119898) 119898(119902

119898)

119899(119902119898) 119897(119902119898) and 119903(119902


intensity value



119894= (1199011

119894 1199012

119894 119901

119889

119894) where 119901119889



119894) where V119889

119894is


119894 1198871199012

119894 119887119901

119889

119894)


(1198921199011 1198921199012 119892119901



119881119889

119894= 119881119889

119894+ 1198881sdot 1199031sdot (119887119901119889

119894minus 119901119889

119894) + 1198882sdot 1199032sdot (119892119901119889minus 119901119889

119894) (13)

119901119889

119894= 119901119889

119894+ 120575119881119889

119894 (14)

In (13)



the range [01]


Fitness calculation


Terminationreached

Updating swarm

Stop

Yes

No



119881119889


119901119889


119887119901119889




1198991198881=2

3(1198881max minus 1198881min) (

119891min119891avg

+119891min2119891max

) + 1198881min

1198991198882=2

3(1198882max minus 1198882min) (

119891min119891avg

+119891min2119891max

) + 1198882min

(15)










fitness = min (119901119889) (119888) max (119897119904) (119887119908) (119901119904) (16)




1and 119899119888

2


119881119889

119894= 119908119889119881119889

119894+ 1198991198881sdot 1199031sdot (119887119901119889

119894minus 119901119889

119894) + 119899119888

2sdot 1199032

sdot (119892119901119889minus 119901119889

119894)

(17)




43 Image Retrieval













119901 =119873119903119903

119873119903

119877 =119873119903119903

119873119889

(19)






) (20)

















(a) (b)


(a) (b)


(a) (b)






0010203040506070809

1


F-measure







6 Conclusions




References



































Volume 2014




Journal of











Journal of


Function Spaces






Algebra















119898)




119866119883=[[

[

minus1 0 1

minus2 0 2

minus1 0 1

]]

]

(2)


119866119884=[[

[

1 2 1

0 0 0


]]

]

(3)


|119866| =1003816100381610038161003816119866119883

1003816100381610038161003816 +1003816100381610038161003816119866119884

1003816100381610038161003816 (4)


120579 = arctan(1003816100381610038161003816119866119883

10038161003816100381610038161003816100381610038161003816119866119884

1003816100381610038161003816

) (5)


110 94 104

98 81 79

73 56 66

1 1 1

1 0

1 0 011100001

255

Image






119898)





LBP (119909119888 119910119888) =

7

sum

119899=0

119904 (119897119899minus 119897119888) 2119899 (6)


(119909119888 119910119888) and 119897



119899)


119899)




119899) and






119896((119894119899)





119901(119909)119894 =

119899119894

119899 0 le 119894 lt 119866 (7)





119899) and 119888(119902

119898) can

be extracted



A B

C D

1 2

3 4





119894119894 (119909 119910) = sum

1199091015840le119909

sum

1199101015840le119910

119894 (1199091015840 1199101015840) (8)


119904 (119909 119910) = 119904 (119909 119910 minus 1) + 119894 (119909 119910)

119894119894 (119909 119910) = 119894119894 (119909 minus 1 119910) + 119904 (119909 119910)

(9)





119899 119910119899) where119910

1=



= 12119898 12119897 for 1199101= 0 1


(iii) For 119905 = 1 119879


119895=1119908119905119895



120576119905= min119891119901120579

sum

119894

119908119894

1003816100381610038161003816ℎ (119909119894 119891 119901 120579) minus 1199101198941003816100381610038161003816 (10)

(3) define ℎ119905(119909) = ℎ(119909 119891

119905 119901119905 120579119905) where 119891

119905 119901119905 and


119894


119908119905+1119894

= 1199081199051198941205731minus119890119894 (11)

where 119890119894= 0 if example 119909



119905= 120576119905(1 minus 120576

119905)


119862 (119909) =

1 if119879

sum

119905=1

120572119905ℎ119905 (119909) ge

1

2

119879

sum

119905=1

120572119905

0 otherwise(12)

where 120572119905= log(1120573

119905)


119898(119894119899) 119899(119894119899) 119897(119894119899) and 119903(119894

119899) and query images 119891(119902

119898) 119898(119902

119898)

119899(119902119898) 119897(119902119898) and 119903(119902


intensity value



119894= (1199011

119894 1199012

119894 119901

119889

119894) where 119901119889



119894) where V119889

119894is


119894 1198871199012

119894 119887119901

119889

119894)


(1198921199011 1198921199012 119892119901



119881119889

119894= 119881119889

119894+ 1198881sdot 1199031sdot (119887119901119889

119894minus 119901119889

119894) + 1198882sdot 1199032sdot (119892119901119889minus 119901119889

119894) (13)

119901119889

119894= 119901119889

119894+ 120575119881119889

119894 (14)

In (13)



the range [01]


Fitness calculation


Terminationreached

Updating swarm

Stop

Yes

No



119881119889


119901119889


119887119901119889




1198991198881=2

3(1198881max minus 1198881min) (

119891min119891avg

+119891min2119891max

) + 1198881min

1198991198882=2

3(1198882max minus 1198882min) (

119891min119891avg

+119891min2119891max

) + 1198882min

(15)










fitness = min (119901119889) (119888) max (119897119904) (119887119908) (119901119904) (16)




1and 119899119888

2


119881119889

119894= 119908119889119881119889

119894+ 1198991198881sdot 1199031sdot (119887119901119889

119894minus 119901119889

119894) + 119899119888

2sdot 1199032

sdot (119892119901119889minus 119901119889

119894)

(17)




43 Image Retrieval













119901 =119873119903119903

119873119903

119877 =119873119903119903

119873119889

(19)






) (20)

















(a) (b)


(a) (b)


(a) (b)






0010203040506070809

1


F-measure







6 Conclusions




References



































Volume 2014




Journal of











Journal of


Function Spaces






Algebra












119899) and






119896((119894119899)





119901(119909)119894 =

119899119894

119899 0 le 119894 lt 119866 (7)





119899) and 119888(119902

119898) can

be extracted



A B

C D

1 2

3 4





119894119894 (119909 119910) = sum

1199091015840le119909

sum

1199101015840le119910

119894 (1199091015840 1199101015840) (8)


119904 (119909 119910) = 119904 (119909 119910 minus 1) + 119894 (119909 119910)

119894119894 (119909 119910) = 119894119894 (119909 minus 1 119910) + 119904 (119909 119910)

(9)





119899 119910119899) where119910

1=



= 12119898 12119897 for 1199101= 0 1


(iii) For 119905 = 1 119879


119895=1119908119905119895



120576119905= min119891119901120579

sum

119894

119908119894

1003816100381610038161003816ℎ (119909119894 119891 119901 120579) minus 1199101198941003816100381610038161003816 (10)

(3) define ℎ119905(119909) = ℎ(119909 119891

119905 119901119905 120579119905) where 119891

119905 119901119905 and


119894


119908119905+1119894

= 1199081199051198941205731minus119890119894 (11)

where 119890119894= 0 if example 119909



119905= 120576119905(1 minus 120576

119905)


119862 (119909) =

1 if119879

sum

119905=1

120572119905ℎ119905 (119909) ge

1

2

119879

sum

119905=1

120572119905

0 otherwise(12)

where 120572119905= log(1120573

119905)


119898(119894119899) 119899(119894119899) 119897(119894119899) and 119903(119894

119899) and query images 119891(119902

119898) 119898(119902

119898)

119899(119902119898) 119897(119902119898) and 119903(119902


intensity value



119894= (1199011

119894 1199012

119894 119901

119889

119894) where 119901119889



119894) where V119889

119894is


119894 1198871199012

119894 119887119901

119889

119894)


(1198921199011 1198921199012 119892119901



119881119889

119894= 119881119889

119894+ 1198881sdot 1199031sdot (119887119901119889

119894minus 119901119889

119894) + 1198882sdot 1199032sdot (119892119901119889minus 119901119889

119894) (13)

119901119889

119894= 119901119889

119894+ 120575119881119889

119894 (14)

In (13)



the range [01]


Fitness calculation


Terminationreached

Updating swarm

Stop

Yes

No



119881119889


119901119889


119887119901119889




1198991198881=2

3(1198881max minus 1198881min) (

119891min119891avg

+119891min2119891max

) + 1198881min

1198991198882=2

3(1198882max minus 1198882min) (

119891min119891avg

+119891min2119891max

) + 1198882min

(15)










fitness = min (119901119889) (119888) max (119897119904) (119887119908) (119901119904) (16)




1and 119899119888

2


119881119889

119894= 119908119889119881119889

119894+ 1198991198881sdot 1199031sdot (119887119901119889

119894minus 119901119889

119894) + 119899119888

2sdot 1199032

sdot (119892119901119889minus 119901119889

119894)

(17)




43 Image Retrieval













119901 =119873119903119903

119873119903

119877 =119873119903119903

119873119889

(19)






) (20)

















(a) (b)


(a) (b)


(a) (b)






0010203040506070809

1


F-measure







6 Conclusions




References



































Volume 2014




Journal of











Journal of


Function Spaces






Algebra











120576119905= min119891119901120579

sum

119894

119908119894

1003816100381610038161003816ℎ (119909119894 119891 119901 120579) minus 1199101198941003816100381610038161003816 (10)

(3) define ℎ119905(119909) = ℎ(119909 119891

119905 119901119905 120579119905) where 119891

119905 119901119905 and


119894


119908119905+1119894

= 1199081199051198941205731minus119890119894 (11)

where 119890119894= 0 if example 119909



119905= 120576119905(1 minus 120576

119905)


119862 (119909) =

1 if119879

sum

119905=1

120572119905ℎ119905 (119909) ge

1

2

119879

sum

119905=1

120572119905

0 otherwise(12)

where 120572119905= log(1120573

119905)


119898(119894119899) 119899(119894119899) 119897(119894119899) and 119903(119894

119899) and query images 119891(119902

119898) 119898(119902

119898)

119899(119902119898) 119897(119902119898) and 119903(119902


intensity value



119894= (1199011

119894 1199012

119894 119901

119889

119894) where 119901119889



119894) where V119889

119894is


119894 1198871199012

119894 119887119901

119889

119894)


(1198921199011 1198921199012 119892119901



119881119889

119894= 119881119889

119894+ 1198881sdot 1199031sdot (119887119901119889

119894minus 119901119889

119894) + 1198882sdot 1199032sdot (119892119901119889minus 119901119889

119894) (13)

119901119889

119894= 119901119889

119894+ 120575119881119889

119894 (14)

In (13)



the range [01]


Fitness calculation


Terminationreached

Updating swarm

Stop

Yes

No



119881119889


119901119889


119887119901119889




1198991198881=2

3(1198881max minus 1198881min) (

119891min119891avg

+119891min2119891max

) + 1198881min

1198991198882=2

3(1198882max minus 1198882min) (

119891min119891avg

+119891min2119891max

) + 1198882min

(15)










fitness = min (119901119889) (119888) max (119897119904) (119887119908) (119901119904) (16)




1and 119899119888

2


119881119889

119894= 119908119889119881119889

119894+ 1198991198881sdot 1199031sdot (119887119901119889

119894minus 119901119889

119894) + 119899119888

2sdot 1199032

sdot (119892119901119889minus 119901119889

119894)

(17)




43 Image Retrieval













119901 =119873119903119903

119873119903

119877 =119873119903119903

119873119889

(19)






) (20)

















(a) (b)


(a) (b)


(a) (b)






0010203040506070809

1


F-measure







6 Conclusions




References



































Volume 2014




Journal of











Journal of


Function Spaces






Algebra













fitness = min (119901119889) (119888) max (119897119904) (119887119908) (119901119904) (16)




1and 119899119888

2


119881119889

119894= 119908119889119881119889

119894+ 1198991198881sdot 1199031sdot (119887119901119889

119894minus 119901119889

119894) + 119899119888

2sdot 1199032

sdot (119892119901119889minus 119901119889

119894)

(17)




43 Image Retrieval













119901 =119873119903119903

119873119903

119877 =119873119903119903

119873119889

(19)






) (20)

















(a) (b)


(a) (b)


(a) (b)






0010203040506070809

1


F-measure







6 Conclusions




References



































Volume 2014




Journal of











Journal of


Function Spaces






Algebra























(a) (b)


(a) (b)


(a) (b)






0010203040506070809

1


F-measure







6 Conclusions




References



































Volume 2014




Journal of











Journal of


Function Spaces






Algebra










0010203040506070809

1


F-measure







6 Conclusions




References



































Volume 2014




Journal of











Journal of


Function Spaces






Algebra





























Volume 2014




Journal of











Journal of


Function Spaces






Algebra









research article semantic-based facial image-retrieval...

Documents